Prove that $$\sqrt{\frac{a^TA^2a}{a^TAa}}\le \sqrt{\|A\|}$$ where $A$ is a $n \times n$ symmetric positive definite matrix and $a \in \mathbb{R}^n \setminus \{0_n\}$.

The norm of a vector here is the Euclidean norm and norm of a matrix $A$ is $$\|A\|=\max\{\|Ax\|:\|x\|=1\}$$ and if $A$ is symmetric then $$\|A\|=\max\{|x^TAx|:\|x\|=1\}$$ and also equal to the largest eigenvalue of $A$.

We have from this min-max theorem that $$\lambda_{min}\|x\|^2\leq x^THx \le \lambda_{max}\|x\|^2$$ where $\lambda_{min},\lambda_{max}$ are the smallest and largest eigenvalues of the Hermitian matrix $H$. So we have

$$\sqrt{\frac{a^TA^2a}{a^TAa}}\le \sqrt{\frac{\lambda^2_{max}}{\lambda_{min}}}$$

In order to prove the inequality in the question, we need to show that $\frac{\lambda^2_{max}}{\lambda_{min}}\le \lambda_{max}$ which implies $\frac{\lambda_{max}}{\lambda_{min}}\le 1$. We have $\|A\|=\lambda_{max}$ and $\|A^{-1}\|=\frac{1}{\lambda_{min}}$ but on the contrary it is a fact that $\|A\|\|A^{-1}\|\ge 1$. Where have I gone wrong and how do we prove this inequality?

  • 1
    $\begingroup$ i think your problem lies in the fact that you should not consider the maximal and minimal eigenvalues, because the vector a is the same in the denominator and the nominator. I'll write out the full solution in a sec $\endgroup$ – Thijs Steel Jan 25 '18 at 11:22
  • $\begingroup$ So you're saying that I've overestimated by taking $\frac{\lambda^2_{max}}{\lambda_{min}}$ right? $\endgroup$ – Picasso Jan 25 '18 at 12:29
  • $\begingroup$ that is indeed what i'm saying $\endgroup$ – Thijs Steel Jan 25 '18 at 12:53

Since $\rm A$ is symmetric and positive definite, it has a symmetric and positive definite square root $\mathrm A^\frac 12$ and $\lambda_{\max} \left( \mathrm A \right) = \| \mathrm A\|_2$, i.e., the spectral radius is equal to the spectral norm. Let $\rm v := \mathrm A^\frac 12 u$. Hence,

$$\dfrac{\mathrm u^\top \mathrm A^2 \mathrm u}{\mathrm u^\top \mathrm A \,\, \mathrm u} = \dfrac{\mathrm u^\top \mathrm A^\frac 12 \mathrm A \, \mathrm A^\frac 12 \mathrm u}{\mathrm u^\top \mathrm A^\frac 12 \mathrm A^\frac 12 \mathrm u} = \dfrac{\mathrm v^\top \mathrm A \, \mathrm v}{\mathrm v^\top \mathrm v} \leq \lambda_{\max} \left( \mathrm A \right) = \| \mathrm A\|_2$$

  • $\begingroup$ We actually need $\frac{u^TA^2u}{u^TAu} \leq \|A\|$. We know from here that there exists a symmetric positive definite square root of $A$, say which is $R$ here. Then we have $\frac{\|R^Tv\|^2}{\|v\|^2} \le \|R^T\|^2 = \|A\|$. $\endgroup$ – Picasso Jan 26 '18 at 11:08
  • $\begingroup$ I don't quite follow. We need $\sqrt{\|A\|}$ according to the problem, so in your proof we need to obtain $\|A\|$ as the upper bound. $\endgroup$ – Picasso Jan 26 '18 at 11:15
  • $\begingroup$ Yes, although the $\lambda$ part is unnecessary because we have for matrix $A$ and vector $v$, $\|Av\| \le \|A\|\|v\|$. So we directly have $\frac{\|A^{1/2}v\|^2}{\|v\|^2} \le \|A^{1/2}\|^2 = \|A\|$. $\endgroup$ – Picasso Jan 26 '18 at 11:36
  • $\begingroup$ I think it is now correct. The proof keeps getting shorter. $\endgroup$ – Rodrigo de Azevedo Jan 26 '18 at 11:48
  • $\begingroup$ We have $\frac{v^TAv}{v^Tv}\le \|A\|$ directly from the definition for norm of a symmetric matrix. Because $\frac{v^TAv}{v^Tv}=\frac{v^TAv}{\|v\|^2}=\frac{v^T}{\|v\|}A\frac{v}{\|v\|} \le \|A\|$. $\endgroup$ – Picasso Jan 26 '18 at 11:52

Let $A = X \Lambda X^{-1}$ be the eigenvalue decomposition for the matrix, this should always exist for a positive definite matrix. Also, $a = \mu_1 v_1 + ...$ denotes the decomposition of a into the eigenvectors of $A$.

Then $\sqrt{\frac{a^TA^2a}{a^TAa}} = \sqrt{\frac{\sum \mu_i^2 \lambda_i^2}{\sum \mu_i^2 \lambda_i}}$.

By calculating the derivative, we find that critical points allow only one $\mu_i$ to be non-zero. This proves that $a$ should be an eigenvector.

Should $a$ be an eigenvector with eigenvalue $\lambda_i$, $\sqrt{\frac{\sum \mu_i^2 \lambda_i^2}{\sum \mu_i^2 \lambda_i}}$ reduces to $\sqrt{\lambda_i}$ which is smaller than the norm of $A$.

  • $\begingroup$ Thanks. Is it possible to provide an answer that doesn't require the derivative? $\endgroup$ – Picasso Jan 25 '18 at 12:34
  • $\begingroup$ i can't think of something that wouldn't require a derivative. Perhaps some theory regarding rational functions. Not my area of expertise $\endgroup$ – Thijs Steel Jan 25 '18 at 12:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.